Scholarlink

<p>Early detection and diagnosis of cognitive impairment are crucial for delaying disease progression and improving clinical management. However, existing approaches that rely on a limited number of EEG channels often suffer from insufficient classification accuracy. To address this limitation, this paper proposes a deep learning framework termed EEG-RCPNet, which integrates attention mechanisms with parallel residual convolutional neural networks to classify cognitively normal and cognitively impaired subjects using two-channel resting-state EEG signals. The proposed architecture first employs a position attention deep convolutional neural network (PAD-CNN) module, in which sequential convolutional operations combined with a position attention mechanism are used to extract shallow multi-scale temporal features from raw EEG signals. Subsequently, a parallel residual channel attention convolutional neural network (PRCA-CNN) module is introduced to capture discriminative deep features across different EEG frequency bands. Finally, a depthwise separable convolutional neural network (DS-CNN) module is applied to refine global feature representations within each branch, and the features extracted from parallel branches are concatenated to enhance multi-scale feature learning capability. The performance of EEG-RCPNet was evaluated using cross-subject ten-fold cross-validation on two datasets: the public ADFTD dataset and a newly constructed CIRS dataset consisting of 212 participants. The proposed model achieved classification accuracies of 0.92 on the ADFTD dataset and 0.86 on the CIRS dataset. On the ADFTD dataset, EEG-RCPNet attained a sensitivity of 0.92, specificity of 0.95, F1 score of 0.93, Macro-F1 score of 0.92, and an MCC of 0.87. On the CIRS dataset, the model maintained a high sensitivity of 0.94 with a specificity of 0.74, yielding an F1 score of 0.90, Macro-F1 score of 0.84, and MCC of 0.71. These results indicate that reliable detection of cognitive impairment might be achieved using limited-channel EEG signals, highlighting the potential of the proposed approach for low-cost, convenient cognitive impairment screening and early-stage monitoring.</p> Graphical abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

EEG-RCPNet: a cross-subject cognitive impairment detection model based on limited-channel eeg signals

  • 期刊论文

  • 发表日期：

    2026年3月25日

查看全文
  • Zining Wang,
  • Minghui Zhao,
  • Keming Cao,
  • Lulu Zhao,
  • Shuo Zhang,
  • Yamei Bai,
  • Jianqing Li,
  • Chengyu Liu

摘要

Early detection and diagnosis of cognitive impairment are crucial for delaying disease progression and improving clinical management. However, existing approaches that rely on a limited number of EEG channels often suffer from insufficient classification accuracy. To address this limitation, this paper proposes a deep learning framework termed EEG-RCPNet, which integrates attention mechanisms with parallel residual convolutional neural networks to classify cognitively normal and cognitively impaired subjects using two-channel resting-state EEG signals. The proposed architecture first employs a position attention deep convolutional neural network (PAD-CNN) module, in which sequential convolutional operations combined with a position attention mechanism are used to extract shallow multi-scale temporal features from raw EEG signals. Subsequently, a parallel residual channel attention convolutional neural network (PRCA-CNN) module is introduced to capture discriminative deep features across different EEG frequency bands. Finally, a depthwise separable convolutional neural network (DS-CNN) module is applied to refine global feature representations within each branch, and the features extracted from parallel branches are concatenated to enhance multi-scale feature learning capability. The performance of EEG-RCPNet was evaluated using cross-subject ten-fold cross-validation on two datasets: the public ADFTD dataset and a newly constructed CIRS dataset consisting of 212 participants. The proposed model achieved classification accuracies of 0.92 on the ADFTD dataset and 0.86 on the CIRS dataset. On the ADFTD dataset, EEG-RCPNet attained a sensitivity of 0.92, specificity of 0.95, F1 score of 0.93, Macro-F1 score of 0.92, and an MCC of 0.87. On the CIRS dataset, the model maintained a high sensitivity of 0.94 with a specificity of 0.74, yielding an F1 score of 0.90, Macro-F1 score of 0.84, and MCC of 0.71. These results indicate that reliable detection of cognitive impairment might be achieved using limited-channel EEG signals, highlighting the potential of the proposed approach for low-cost, convenient cognitive impairment screening and early-stage monitoring.

Graphical abstract